Research On ECG Biometrics Based On Metric Learning

As artificial intelligence theory and technology increasingly mature,the production and the life styles in the society are constantly changing.Biometric identification technology has replaced the traditional identification methods in some scenarios,bringing to the human life a big convenience.As one of the most widely used medical signals,ECG signal is also applied to biometrics.Compared with fingerprint,face recognition and other biometric technologies,ECG biometrics has many advantages such as liveness detection and low risk of data exposure.Therefore,ECG biometrics has attracted much attention.Although the researches of ECG biometrics have made some progress in recent years,there are still many challenging problems in this field.These problems are concentrated mainly on three aspects:the first one is the dynamic morphological variability of homologous ECG signals caused by human emotion,disease and other factors;the second is the low signal-tonoise ratio caused by noise such as power-line interference,myoelectric interference and baseline drift in the process of acquisition;the third is more unstable factors caused by portable acquisition such as finger acquisition.All of the above issues lead to the problem that the distance between homologous samples is larger than that between heterogeneous samples(i.e.the problem of intra-class variation),which is essentially a metric problem.The application of metric learning to extracting features with large distance between classes and small distance within class has achieved excellent performance in the fields like face recognition and pedestrian re-identification.Therefore,this thesis proposes the following two ECG biometrics methods based on metric learning to solve the above problems.1.Kernel marginal Fisher analysis for ECG biometrics.In this thesis,metric learning is used for ECG biometrics to solve the above problems,and ECG biometrics method based on kernel marginal Fisher analysis is proposed.Firstly,ECG signals are implicitly mapped into higher-dimensional feature space by kernel function to solve the problem which are difficult to divide linearly in original space.Then,by using marginal Fisher analysis in high-dimensional space,the features are extracted by minimizing intra-class distance and maximizing inter-class distance.Finally,the effectiveness of the method is evaluated on two clinical ECG databases MITDB and PTBDB.2.Dual-domain low-rank fusion deep metric learning for ECG biometrics.Capturing ECG signals from fingertips is one of the new trends in ECG biometrics.However,compared with clinical ECG signal,the finger ECG signal has more serious noises and dynamic morphological variability.In this thesis,dual-domain low-rank fusion deep metric learning for finger ECG biometrics is proposed.To learn intra-individual compact features,AAM-Loss(Additive Angular Margin Loss)is introduced to control training.To enrich the discriminant information of features,time domain features and frequency domain features are combined.Low-rank fusion method is also adopted to not only preserve the context information of features,but also mine the relationship between time domain features and frequency domain features.In addition,the proposed method is robust enough to bad quality finger ECG signals and does not require noise and outliers removal.Finally,experiments on finger ECG databases UOFTDB and CYBHiDB demonstrate that the proposed method outperforms the state-of-theart methods.Additionally,ablation experiments show the effectiveness of every part of our framework.